Re: [RFC PATCH 1/8] fs: introduce get_shared_files() for dax&reflink

From: Ruan Shiyang
Date: Mon Aug 10 2020 - 04:07:30 EST




On 2020/8/8 上午12:15, Darrick J. Wong wrote:
On Fri, Aug 07, 2020 at 09:13:29PM +0800, Shiyang Ruan wrote:
Under the mode of both dax and reflink on, one page may be shared by
multiple files and offsets. In order to track them in memory-failure or
other cases, we introduce this function by finding out who is sharing
this block(the page) in a filesystem. It returns a list that contains
all the owners, and the offset in each owner.

For XFS, rmapbt is used to find out the owners of one block. So, it
should be turned on when we want to use dax&reflink feature together.

Signed-off-by: Shiyang Ruan <ruansy.fnst@xxxxxxxxxxxxxx>
---
fs/xfs/xfs_super.c | 67 +++++++++++++++++++++++++++++++++++++++++++++
include/linux/dax.h | 7 +++++
include/linux/fs.h | 2 ++
3 files changed, 76 insertions(+)

diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c
index 379cbff438bc..b71392219c91 100644
--- a/fs/xfs/xfs_super.c
+++ b/fs/xfs/xfs_super.c
@@ -35,6 +35,9 @@
#include "xfs_refcount_item.h"
#include "xfs_bmap_item.h"
#include "xfs_reflink.h"
+#include "xfs_alloc.h"
+#include "xfs_rmap.h"
+#include "xfs_rmap_btree.h"
#include <linux/magic.h>
#include <linux/fs_context.h>
@@ -1097,6 +1100,69 @@ xfs_fs_free_cached_objects(
return xfs_reclaim_inodes_nr(XFS_M(sb), sc->nr_to_scan);
}
+static int _get_shared_files_fn(

Needs an xfs_ prefix...

+ struct xfs_btree_cur *cur,
+ struct xfs_rmap_irec *rec,
+ void *priv)
+{
+ struct list_head *list = priv;
+ struct xfs_inode *ip;
+ struct shared_files *sfp;
+
+ /* Get files that incore, filter out others that are not in use. */
+ xfs_iget(cur->bc_mp, cur->bc_tp, rec->rm_owner, XFS_IGET_INCORE, 0, &ip);

No error checking at all?

What if rm_owner refers to metadata?

Yes, we need to check whether the page contains metadata. I remembered that. I wrote the check code but removed it in this patch, because I didn't find a way to associate this block device with the dax page that contains metadata. We can call dax_associate_entry() to create the association if the page's owner is a file, but it's not work for metadata. I should have explained here.


+ if (ip && !ip->i_vnode.i_mapping)
+ return 0;

When is the xfs_inode released? We don't iput it here, and there's no
way for dax_unlock_page (afaict the only consumer) to do it, so we
leak the reference.

Do you mean xfs_irele() ? Sorry, I didn't realize that. All we need is get the ->mapping form a given inode number. So, I think we can call xfs_irele() when exiting this function.


+
+ sfp = kmalloc(sizeof(*sfp), GFP_KERNEL);

If there are millions of open files reflinked to this range of pmem this
is going to allocate a /lot/ of memory.

+ sfp->mapping = ip->i_vnode.i_mapping;

sfp->mapping = VFS_I(ip)->i_mapping;

+ sfp->index = rec->rm_offset;
+ list_add_tail(&sfp->list, list);

Why do we leave ->cookie uninitialized? What does it even do?

It's my fault. ->cookie should have been added in the next patch. It stores each owner's dax entry when calling dax_lock_page() in memory-failure.


+
+ return 0;
+}
+
+static int
+xfs_fs_get_shared_files(
+ struct super_block *sb,
+ pgoff_t offset,

Which device does this offset refer to? XFS supports multiple storage
devices.

Also, uh, is this really a pgoff_t? If yes, you can't use it with
XFS_B_TO_FSB below without first converting it to a loff_t.

The offset here is assigned as iomap->addr, which is obtained from iomap_begin(). So that we can easily looking up for the owners of this offset.

I don't quite understand what you said about supporting multiple storage devices. What are these devices? Do you mean NVDIMM, HDD, SSD and others?


+ struct list_head *list)
+{
+ struct xfs_mount *mp = XFS_M(sb);
+ struct xfs_trans *tp = NULL;
+ struct xfs_btree_cur *cur = NULL;
+ struct xfs_rmap_irec rmap_low = { 0 }, rmap_high = { 0 };

No need to memset(0) rmap_low later, or zero rmap_high just to memset it
later.

+ struct xfs_buf *agf_bp = NULL;
+ xfs_agblock_t bno = XFS_B_TO_FSB(mp, offset);

"FSB" refers to xfs_fsblock_t. You just ripped the upper 32 bits off
the fsblock number.

I think I misused these types. Sorry for that.

+ xfs_agnumber_t agno = XFS_FSB_TO_AGNO(mp, bno);
+ int error = 0;
+
+ error = xfs_trans_alloc_empty(mp, &tp);
+ if (error)
+ return error;
+
+ error = xfs_alloc_read_agf(mp, tp, agno, 0, &agf_bp);
+ if (error)
+ return error;
+
+ cur = xfs_rmapbt_init_cursor(mp, tp, agf_bp, agno);
+
+ memset(&cur->bc_rec, 0, sizeof(cur->bc_rec));

Not necessary, bc_rec is zero in a freshly created cursor.

+ /* Construct the range for one rmap search */
+ memset(&rmap_low, 0, sizeof(rmap_low));
+ memset(&rmap_high, 0xFF, sizeof(rmap_high));
+ rmap_low.rm_startblock = rmap_high.rm_startblock = bno;
+
+ error = xfs_rmap_query_range(cur, &rmap_low, &rmap_high,
+ _get_shared_files_fn, list);
+ if (error == -ECANCELED)
+ error = 0;
+
+ xfs_btree_del_cursor(cur, error);
+ xfs_trans_brelse(tp, agf_bp);
+ return error;
+}

Looking at this, I don't think this is the right way to approach memory
poisoning. Rather than allocating a (potentially huge) linked list and
passing it to the memory poison code to unmap pages, kill processes, and
free the list, I think:

1) "->get_shared_files" should be more targetted. Call it ->storage_lost
or something, so that it only has one purpose, which is to react to
asynchronous notifications that storage has been lost.

Yes, it's better. I only considered file tracking.

2) The inner _get_shared_files_fn should directly call back into the
memory manager to remove a poisoned page from the mapping and signal
whatever process might have it mapped.

For the error handling part, it's really reasonable. But we have to call dax_lock_page() at the beginning of memory-failure, which also need to iterate all the owners in order to lock their dax entries. I think it not supposed to call the lock in each iteration instead of at the beginning.


That way, _get_shared_files_fn can look in the xfs buffer cache to see
if we have a copy in DRAM, and immediately write it back to pmem.

I didn't think of fixing the storage device. Will take it into consideration.


Hmm and now that you've gotten me rambling about hwpoison, I wonder what
happens if dram backing part of the xfs buffer cache goes bad...

Yes, so many possible situations to consider. For the current stage, just shutdown the filesystem if memory failures on metadata, and kill user processes if failures on normal files. Is that OK?


Anyway, thanks for reviewing.

--
Thanks,
Ruan Shiyang.


--D

+
static const struct super_operations xfs_super_operations = {
.alloc_inode = xfs_fs_alloc_inode,
.destroy_inode = xfs_fs_destroy_inode,
@@ -1110,6 +1176,7 @@ static const struct super_operations xfs_super_operations = {
.show_options = xfs_fs_show_options,
.nr_cached_objects = xfs_fs_nr_cached_objects,
.free_cached_objects = xfs_fs_free_cached_objects,
+ .get_shared_files = xfs_fs_get_shared_files,
};
static int
diff --git a/include/linux/dax.h b/include/linux/dax.h
index 6904d4e0b2e0..0a85e321d6b4 100644
--- a/include/linux/dax.h
+++ b/include/linux/dax.h
@@ -40,6 +40,13 @@ struct dax_operations {
extern struct attribute_group dax_attribute_group;
+struct shared_files {
+ struct list_head list;
+ struct address_space *mapping;
+ pgoff_t index;
+ dax_entry_t cookie;
+};
+
#if IS_ENABLED(CONFIG_DAX)
struct dax_device *dax_get_by_host(const char *host);
struct dax_device *alloc_dax(void *private, const char *host,
diff --git a/include/linux/fs.h b/include/linux/fs.h
index f5abba86107d..81de3d2739b9 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -1977,6 +1977,8 @@ struct super_operations {
struct shrink_control *);
long (*free_cached_objects)(struct super_block *,
struct shrink_control *);
+ int (*get_shared_files)(struct super_block *sb, pgoff_t offset,
+ struct list_head *list);
};
/*
--
2.27.0